Hi Naveen,

In my limited understanding, SMOTE is just used to ensure that we have "richer" data (specifically, class 1 records) for statistical model to be trained well, and hence, perform well later.

Remember that log-loss score of a statistical model just by itself has very little significance. It is only useful in comparative context. We can compare the log-loss score of two (or more) statistical models and conclude the best performing one, provided they were run on the same dataset i.e. the underlying data (and of course, its distribution) should be constant.

Also, even the best performing model should beat the baseline score (of the naive model) on the same dataset.